[Intel-gfx] [PATCH v2 02/20] drm: Don't update plane properties for atomic planes if it stays the same

[Maarten Lankhorst]

Op 13-07-15 om 11:45 schreef Daniel Vetter:
> On Mon, Jul 13, 2015 at 11:23:45AM +0200, Maarten Lankhorst wrote:
>> Op 13-07-15 om 11:13 schreef Daniel Vetter:
>>> On Mon, Jul 13, 2015 at 10:59:32AM +0200, Maarten Lankhorst wrote:
>>>> Op 08-07-15 om 22:12 schreef Daniel Vetter:
>>>>> On Wed, Jul 08, 2015 at 08:25:07PM +0200, Maarten Lankhorst wrote:
>>>>>> Op 08-07-15 om 19:52 schreef Daniel Vetter:
>>>>>>> On Wed, Jul 08, 2015 at 06:35:47PM +0200, Maarten Lankhorst wrote:
>>>>>>>> Op 08-07-15 om 10:55 schreef Daniel Vetter:
>>>>>>>>> On Wed, Jul 08, 2015 at 10:00:22AM +0200, Maarten Lankhorst wrote:
>>>>>>>>>> Op 07-07-15 om 18:43 schreef Daniel Vetter:
>>>>>>>>>>> On Tue, Jul 07, 2015 at 05:08:34PM +0200, Maarten Lankhorst wrote:
>>>>>>>>>>>> Op 07-07-15 om 14:10 schreef Daniel Vetter:
>>>>>>>>>>>>> On Tue, Jul 07, 2015 at 12:20:10PM +0200, Maarten Lankhorst wrote:
>>>>>>>>>>>>>> Op 07-07-15 om 11:18 schreef Daniel Vetter:
>>>>>>>>>>>>>>> On Tue, Jul 07, 2015 at 09:08:13AM +0200, Maarten Lankhorst wrote:
>>>>>>>>>>>>>>>> This allows the first atomic call during hw init to be a real modeset,
>>>>>>>>>>>>>>>> which is useful for forcing a recalculation.
>>>>>>>>>>>>>>> fbcon is optional, you can't rely on anything being done in any specific
>>>>>>>>>>>>>>> way. What exactly do you need this for, what's the implications?
>>>>>>>>>>>>>> In the hw readout I noticed some warnings when I wasn't setting any mode property in the readout.
>>>>>>>>>>>>>> I want the first function to be the modeset, so we have a sane base to commit changes on.
>>>>>>>>>>>>>> Ideally this whole function would have a atomic counterpart which does it in one go. :)
>>>>>>>>>>>>> Yeah. Otoh as soon as we have atomic modeset working we can replace all
>>>>>>>>>>>>> the legacy entry points with atomic helpers, and then even plane_disable
>>>>>>>>>>>>> will be a full atomic modeset.
>>>>>>>>>>>>>
>>>>>>>>>>>>> What did fall apart with just touching properties/planes now?
>>>>>>>>>>>> Also when i915 is fully atomic it calculates in intel_modeset_compute_config
>>>>>>>>>>>> if a modeset is needed after the first atomic call. Right now because
>>>>>>>>>>>> intel_modeset_compute_config is only called in set_config so this works as expected.
>>>>>>>>>>>> Otherwise drm_plane_force_disable or rotate_0 will force a modeset,
>>>>>>>>>>>> and if the final mode is different this will introduce a double modeset.
>>>>>>>>>>> For expensive properties (i.e. a no-op changes causes something that takes
>>>>>>>>>>> time like modeset or vblank wait) we need to make sure we filter them out
>>>>>>>>>>> in atomic_check. Yeah not quite there yet with pure atomic, but meanwhile
>>>>>>>>>>> the existing legacy set_prop functions should all filter out no-op changes
>>>>>>>>>>> themselves. If we don't do that for rotation then that's a bug.
>>>>>>>>>>>
>>>>>>>>>>> Same for disabling planes harder, that shouldn't take time. Especially
>>>>>>>>>>> since fbcon only force-disable non-primary plane, and for driver load
>>>>>>>>>>> that's the exact thing we already do in the driver anyway.
>>>>>>>>>> Something like this?
>>>>>>>>>> ---
>>>>>>>>>> diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c
>>>>>>>>>> index a1d4e13f3908..2989232f4996 100644
>>>>>>>>>> --- a/drivers/gpu/drm/drm_atomic_helper.c
>>>>>>>>>> +++ b/drivers/gpu/drm/drm_atomic_helper.c
>>>>>>>>>> @@ -30,6 +30,7 @@
>>>>>>>>>>  #include <drm/drm_plane_helper.h>
>>>>>>>>>>  #include <drm/drm_crtc_helper.h>
>>>>>>>>>>  #include <drm/drm_atomic_helper.h>
>>>>>>>>>> +#include "drm_crtc_internal.h"
>>>>>>>>>>  #include <linux/fence.h>
>>>>>>>>>>  
>>>>>>>>>>  /**
>>>>>>>>>> @@ -1716,7 +1717,12 @@ drm_atomic_helper_crtc_set_property(struct drm_crtc *crtc,
>>>>>>>>>>  {
>>>>>>>>>>  	struct drm_atomic_state *state;
>>>>>>>>>>  	struct drm_crtc_state *crtc_state;
>>>>>>>>>> -	int ret = 0;
>>>>>>>>>> +	uint64_t retval;
>>>>>>>>>> +	int ret;
>>>>>>>>>> +
>>>>>>>>>> +	ret = drm_atomic_get_property(&crtc->base, property, &retval);
>>>>>>>>>> +	if (!ret && val == retval)
>>>>>>>>>> +		return 0;
>>>>>>>>>>  
>>>>>>>>>>  	state = drm_atomic_state_alloc(crtc->dev);
>>>>>>>>>>  	if (!state)
>>>>>>>>>> @@ -1776,7 +1782,12 @@ drm_atomic_helper_plane_set_property(struct drm_plane *plane,
>>>>>>>>>>  {
>>>>>>>>>>  	struct drm_atomic_state *state;
>>>>>>>>>>  	struct drm_plane_state *plane_state;
>>>>>>>>>> -	int ret = 0;
>>>>>>>>>> +	uint64_t retval;
>>>>>>>>>> +	int ret;
>>>>>>>>>> +
>>>>>>>>>> +	ret = drm_atomic_get_property(&plane->base, property, &retval);
>>>>>>>>>> +	if (!ret && val == retval)
>>>>>>>>>> +		return 0;
>>>>>>>>>>  
>>>>>>>>>>  	state = drm_atomic_state_alloc(plane->dev);
>>>>>>>>>>  	if (!state)
>>>>>>>>>> @@ -1836,7 +1847,12 @@ drm_atomic_helper_connector_set_property(struct drm_connector *connector,
>>>>>>>>>>  {
>>>>>>>>>>  	struct drm_atomic_state *state;
>>>>>>>>>>  	struct drm_connector_state *connector_state;
>>>>>>>>>> -	int ret = 0;
>>>>>>>>>> +	uint64_t retval;
>>>>>>>>>> +	int ret;
>>>>>>>>>> +
>>>>>>>>>> +	ret = drm_atomic_get_property(&connector->base, property, &retval);
>>>>>>>>>> +	if (!ret && val == retval)
>>>>>>>>>> +		return 0;
>>>>>>>>>>  
>>>>>>>>>>  	state = drm_atomic_state_alloc(connector->dev);
>>>>>>>>>>  	if (!state)
>>>>>>>>> The reason I didn't do this is that a prop change might still result in no
>>>>>>>>> hw state change (e.g. if you go automitic->explicit setting matching
>>>>>>>>> automatic one). Hence I think we need to solve this in lower levels
>>>>>>>>> anyway, i.e. in when computing the config. But it shouldn't cause trouble
>>>>>>>>> yet.
>>>>>>>> Is that a ack or nack?
>>>>>>> I think we shouldn't need this really for i915, and it might cover up
>>>>>>> bugs. I prefer we just do the evade modeset logic you've implemented once
>>>>>>> we switch over to atomic props. Since atm we only have atomic props which
>>>>>>> get updated in pageflips we shouldn't have serious problems here yet (for
>>>>>>> setting the rotation prop to 0° again when fbdev starts up).
>>>>>>>
>>>>>>> Or do I miss something still here?
>>>>>> Yes, if the hardware mode is incompatible with its calculated sw mode,
>>>>>> and we set a different mode from fbdev you get 2 modesets instead of 1.
>>>>> How does that happen? For setting the rotation property we should just
>>>>> duplicate the current crtc state. Since there's no mode changing (they
>>>>> should match perfectly no matter how botched the reconstruction is) there
>>>>> shouldn't be any need to recompute the config completely and discover that
>>>>> there's a mismatch. Which means we'll just do the plane update (which
>>>>> might do a few silly mmios but shouldn't block) and that's it.
>>>>>
>>>>> At least that's what I'd expect - where does this fall apart?
>>>> If crtc is active and primary fb visible, and converted to atomic:
>>>>
>>>> restore_fbdev_mode() ->
>>>> 	drm_mode_plane_set_obj_prop() ->
>>>> 		drm_atomic_helper_plane_set_property() ->
>>>> 			drm_atomic_get_plane_state() ->
>>>> 				drm_atomic_get_crtc_state()
>>>> crtc state is part of the state, intel_modeset_pipe_config performs
>>>> the initial check if modeset's needed. Lets assume yes:
>>> "Let's assume yes" -> that's imo a bug, so where does this happen so that
>>> we can fix it? Disabling a plane or setting a plane prop really shouldn't
>>> result in a modeset. Well at least if it's not a plane prop that does
>>> required a modeset (but I don't think we have any of those).
>> From a driver point of view you wouldn't be able to distinguish it from a real modeset to the same mode. :(
>> In both cases you have all planes added and the crtc.
>>
>> Thinking about it more there will be 1 thing saving us from a modeset,
>> drm_atomic_crtc_check will reject enable without mode_blob for atomic drivers,
>> so until the first mode is set all atomic updates to the crtc will be rejected.
>>
>> Unfortunately you will still get WARN_ON's for this, so a better solution's needed.
> Ok I think I start to grasp what's wrong, the trouble is that we don't
> have the mode stuff fully set up yet (which is part of fastboot), which
> means we'll get a bogus crtc_state->mode_changed despite that nothing
> really changed. Ugly.
No, mode_changed would be harmless, with proper skip modeset support it can be converted to a noop.
> Could we insert a dummy mode_blob to avoid the WARNs and the bogus
> mode_changed instead? The problem really is that doing this here is just
> plugging the one source of troubles you're seeing right now (fbcon), the
> initial set_* calls could come from anything really in any order. So we
> really better be able to cope.
Doesn't this mean we should set a real mode read out from hw state instead?

> Even converting fbdev to have an optional atomic patch for DRIVER_ATOMIC
> (unsafe i915 options aren't a concern here for me) won't fix this.